Explainable Used Car Pricing: A Machine Learning Framework With Analysis of Marginal Effects
研究提出一个三阶段机器学习框架，用于透明地预测二手车价格，并分析二元变量的边际效应，发现车龄、品牌和发动机排量是主要定价因素，集成树算法表现最佳。
Used car pricing is always complex and involves various factors. It is difficult for consumers to obtain full transparency into used car pricing, which sellers often obscure to protect margins. Even if the pricing model for one specific used car can be established, the wide variety of used cars and features further complicates the problem. Motivated by a real-world application in insurance premium pricing for used cars, our study presents a transparently explainable solution to predict various used car prices with elaborate feature engineering. We propose a novel three-phase approach, with each phase dedicated to prediction models, explainable machine learning, and the examination of marginal effects of binary variables. We conduct a comparative analysis to assess the performance of different predicting models with real-world data for robustness. The consistent results suggest that ensemble tree algorithms perform the best, with a mean absolute percentage error of 0.269 (out-of-sample), compared with 0.957 from the linear benchmark. The performance of deep neural networks with numerical features is salient when dealing with more homogeneous and specific data. By examining the individual feature attributes, the car's age, make (i.e., brand), and engine capacity are the top three determinants of the price with strong monotonicity. Finally, the marginal effects of significant binary variables and their interactions on prices are quantified. Our analysis reveals the non-linearity and complexity in effect sizes. The proposed three-phase approach provides a transparent and reliable pricing pipeline for the used car market and can essentially assist stakeholders in making decisions.